Is AI the New China Shock and SAP Is Walling Itself In

Is AI the New China Shock?

The main auditorium at MIT was sold out this week: economists discussed whether generative AI will trigger the same structural disruption as China's entry into the WTO in 2001. David Autor presented data showing that the China Shock cost 2 million US manufacturing jobs—primarily in Rust Belt regions with no service-sector alternatives. The new thesis: This time, AI isn't hitting factory workers, but the knowledge-based middle class. UBS analysts speak of 300 million white-collar jobs at risk worldwide. The crucial difference is speed: China took 20 years to penetrate global markets; GPT-4 reached 100 million users in two months. Ford CEO Jim Farley warned of a 'brutal disruption' in the design and engineering departments. The pattern is familiar: first, simple tasks disappear, then entire functions migrate. → Benedict Evans

Synthszr Take: The AI disruption works like an inverted container port—instead of shipping physical goods more efficiently, algorithms dock directly with knowledge production, making transport obsolete. Malcolm McLean revolutionized freight shipping in 1956 by introducing standardized containers; today, GPT standardizes cognitive work into token packages. The China Shock still needed factories, ships, and ports as infrastructure. AI only needs API access. The geographic buffer zone is disappearing: A prompt in Bangalore competes in real-time with an analyst in Manhattan. What economists overlook: Unlike Chinese factories, AI models improve exponentially with use. This is no longer a trade shock, but a self-accelerating wave of substitution.

Chinese AI Companies Announce Massive Funding Rounds

Within a few days, three Chinese AI companies announced massive funding rounds. Moonshot AI closed a $2 billion round at a valuation of over $20 billion, led by Meituan. Moonshot AI reports a milestone: its new language model, Kimi K2.5, generated more revenue in the first 20 days after launch than in all of 2025. StepFun is currently structuring for a potential Hong Kong IPO and is simultaneously raising $2.5 billion. DeepSeek is negotiating a funding round that would value the company at $45 to $50 billion—potential lead investors include China's state-backed semiconductor investment fund. The logic behind the timing: With Zhipu ($52 billion market cap) and MiniMax ($32 billion), there have been publicly traded comparables in Hong Kong since early May. Private valuations are no longer floating in a vacuum. DeepSeek's round is particularly explosive: The national AI fund, with a volume of 60 billion RMB, could invest up to $7.35 billion, with founder Liang Wenfeng himself intending to make the largest single contribution. → Hello China Tech

Synthszr Take: DeepSeek's financing operates like China's semiconductor catch-up strategy after the 2022 US sanctions: The Big Fund, originally intended for chip fabs, is now investing in AI companies that prove hardware limitations can be compensated by algorithmic efficiency. As SMIC demonstrates with 7nm chips without EUV lithography, the second path is often more innovative than the first—the forced resource scarcity became a driver of innovation. The $45 billion valuation is less about market mechanics and more of a political signal: Beijing is building its own technology champions, independent of Western hardware and capital. The Big Fund is financing not just chips, but an entire ecosystem. DeepSeek's open-source strategy isn't altruism; it's a geopolitical offensive.

In a Panic, SAP Walls Itself In and Locks Out Agents

SAP is acquiring the Freiburg-based startup Prior Labs for an undisclosed amount paid 'almost entirely in cash' and plans to invest one billion euros over four years in a European AI lab for structured enterprise data. The lab will specialize in tabular foundation models—precisely the data type that dominates SAP systems. Simultaneously, the Walldorf-based company is radically updating its API policies: all third-party agents like OpenClaw will be locked out, with only SAP-approved systems like its in-house Joule or Nvidia's NemoClaw still granted access. The message is unmistakable: SAP wants to maintain control over all AI access to its customer data. The timing is no coincidence—it comes just as agent startups are beginning to challenge the decades-long data dominance of ERP giants. → TheSequence

Synthszr Take: SAP's strategy is reminiscent of German city zoning, where preservation triumphs over development. Just as municipalities use rigid building codes to prevent modern architecture from challenging historic structures, SAP is erecting a regulatory wall around its data treasures—only approved 'builders' are allowed in, while the rest are kept out. The billion for Prior Labs is less about innovation and more about deepening the moat. SAP isn't buying the future; it's buying time. The problem: Agent technology evolves faster than zoning plans can be changed. Those who build walls today risk watching from their own fortress tomorrow as value creation happens outside.

HubSpot Reinvents Performance Pricing

Starting April 14, HubSpot is introducing a radically new pricing model for its AI agents: customers will only pay upon successful completion of a task. The Breeze Customer Agent will cost $0.50 per resolved conversation (instead of $1 per conversation), and the Prospecting Agent will cost $1 per qualified lead (instead of a monthly flat fee). The company justifies the move with the superior performance of its agents: The Customer Agent already resolves 65% of all conversations and reduces handling time by 39%. Over 8,000 customers are actively using the tool, and activations for the Prospecting Agent recently grew by 57% quarter-over-quarter. According to Chief Customer Officer Jon Dick, only agents with access to the full CRM context can deliver such consistent results—generic AI tools without enterprise context cannot achieve this level of reliability. → Evan Armstrong from The Leverage

Synthszr Take: HubSpot's performance pricing works like piece-rate pay in 20th-century manufacturing—only here, the machine itself is paid per piece. On the factory floor, the rule was: produce more parts, earn more; deliver defects, get nothing. HubSpot applies this principle to AI agents, making two bets: first, that its 65% success rate will remain stable. Second, that customers are willing to pay more for solved problems than for attempts. The real innovation isn't the pricing, but the self-confidence: only a company that sees its CRM context as an unbeatable advantage can afford such a model. For the enterprise AI economy, this means the battle between platforms and tools will be decided by data sovereignty, not by features.

Agents That Build Their Own Infrastructure

The next leap in AI development isn't a better model—it's agents that rewrite their own code. Sakana AI's Darwin-Gödel Machine increased its performance on the SWE benchmark from 20% to 50% by independently introducing validation steps and improving its file-viewing functions. Meta went even further with Hyperagents: they merge a task agent and a meta-agent into a single, self-editing program. The systems unsolicitedly develop their own memory layers, performance tracking, and multi-stage evaluation pipelines. In a paper-review task, an initially blank agent improved its accuracy from 0.0 to 0.710. The problem: These self-modifying agents break the classic harness architecture that enterprise platforms rely on. → AlphaSignal

Synthszr Take: Self-modifying agents work like a biotech firm rewriting its own lab protocols mid-study—a nightmare for any audit committee, but exactly what groundbreaking research sometimes requires. In the pharma world, the FDA would halt the study immediately: changing protocols during Phase 3 is a sure path to non-approval. Meta's Hyperagents do exactly that: they change not only their task logic but also how they evaluate and improve themselves. Enterprise platforms face a dilemma—either they open their systems to this kind of meta-evolution, or they will be overtaken by native tools that don't have a compliance handbrake. The irony: The more sophisticated the safety mechanisms, the greater the competitive disadvantage against systems that are allowed to reinvent themselves.

Hermes Establishes Itself as an Alternative to OpenClaw

A new movement is emerging: Hermes Agent v0.13.0 has been released, while Reddit polls suggest about 30 percent of OpenClaw users have already switched. The open-source agent from Nous Research has collected over 135,000 GitHub stars since February 2026 and can run on either a $5 VPS or a GPU cluster. The MIT-licensed project comes with 40 pre-installed skills and works with all major language models, from OpenAI and Anthropic to Kimi. Yesterday's version 0.13.0 ('The Tenacity Release') included 864 commits from 295 developers and closed eight critical security vulnerabilities. The kicker: Hermes learns autonomously from complex tasks, extracts reusable patterns, and writes its own skill files for similar future tasks. → The Neuron

Synthszr Take: Hermes operates like the four-eyes principle in auditing, where a junior auditor takes their own review notes for the next period. While OpenClaw was designed as a central message hub, Hermes centers on the learning cycle: after each complex task, the system analyzes what worked and documents the solution as a reusable skill. These self-written work instructions are automatically retrieved for the next similar case, instead of starting from scratch every time. The switch by 30 percent of users shows that if a market leader doesn't question its architecture, an open-source project with a better foundational structure can quickly gain market share. The real innovation isn't the technology, but the fact that an agent builds its own methodological competence.

Google is Chasing Oura

Google is radically transforming Fitbit: The new Fitbit Air comes without a display, joining the ranks of Oura and Whoop. After the 2021 acquisition, Google killed its smartwatch lines to push the Pixel Watch—now it's pivoting to screenless wearables. The strategy is clear: less distraction, more background data collection. In parallel, new hardware players are entering the market: SpeakON attaches magnetically to a smartphone and edits thoughts in real-time, INSPEC tracks dreams for lucid dreaming, and AWEAR measures brainwaves directly in the ear. The wearables market is fragmenting into specialized niches—away from the all-in-one device on the wrist. Google is positioning Fitbit as a data collector without the display stress, while startups experiment with exotic use cases. → Product Hunt Weekly

Synthszr Take: Google is treating Fitbit like a regulatory agency for an orphan drug—the original indication is abandoned in search of a profitable niche in a smaller but more lucrative market. Just as with rare diseases, where the FDA lures companies with fast-track procedures and market exclusivity, Fitbit is finding its second chance in the premium segment of display-free trackers. The irony: Google killed Fitbit's smartwatches for the Pixel Watch, only to now discover that the true value lies in the absence of features. The wearables market is currently undergoing its own version of pharma specialization—every device needs its narrowly defined indication, from dream tracking to brainwave measurement. Google is betting that less screen means more margin.

The Hyperscaler Capex Map: May 2026 Edition

The four major hyperscalers—Microsoft, Amazon, Alphabet, and Meta—reported their quarterly earnings within a 48-hour window at the end of April 2026 and unanimously raised their investment forecasts. With a combined $725 billion for 2026—a 77 percent increase from the previous year's $410 billion—they initiated the largest concentrated infrastructure cycle in technology history. Apple, on the other hand, lowered its capital expenditures and broke with its decades-old 'net cash neutral' policy. A week later, The Information revealed the details of Anthropic's Google Cloud contract: $200 billion over five years for 5 gigawatts of capacity. The four largest US cloud providers now collectively hold about $2.1 trillion in revenue backlog, with half of it—$1.05 trillion—coming from two cash-burning AI startups: OpenAI and Anthropic. The total investment cycle for 2026 amounts to over a trillion dollars when you add Oracle ($50B), Apple ($13B), Neoclouds ($60B), China ($80B), state actors ($60B), and others ($48B). → The Business Engineer

Synthszr Take: The hyperscaler investments follow the pattern of a reinsurance bubble before a natural disaster: primary insurers (OpenAI, Anthropic) write massive policies without a sufficient capital base, while reinsurers (Microsoft, Google) take on the risks, assuming the catastrophe won't happen. In actuarial science, this is called 'spiraling'—when reinsurers have to finance their own clients so they can pay the premiums, a self-reinforcing dependency is created. Half of the trillion-dollar capex is 'guaranteed' by companies operating with negative cash flow. Unlike classic infrastructure cycles, where creditworthy customers justify the investment, here the infrastructure is financing its own demand. This is no longer a tech investment but a macroeconomic bet on the stability of government bonds—just without their collateral.

25 Percent of Top Substack is AI-Generated—And the Trend is Unstoppable

Evan Armstrong analyzed 3,229 posts from the most successful Substack newsletters and found an inconvenient truth: a quarter of business publications consist of AI-generated text. Two of the top 10 newsletters are completely synthetic—no human pen involved. Armstrong, a self-professed hater of AI-generated text, wanted to analytically substantiate his dislike. Instead, the data forced him to question his own thesis. The implication is brutal: If even paying Substack readers don't notice the difference or don't care, what does that mean for any business model based on 'we'll tell you something new'? Armstrong sees his own medium as under threat. → Evan Armstrong from The Leverage

Synthszr Take: The Substack data functions like a blind test in pharmaceutical research—except here, nobody knew a test was even being conducted. In clinical trials, the placebo and the active substance must be indistinguishable, or the test is worthless. On Substack, AI newsletters have achieved this indistinguishability without readers being informed. Paying subscribers consume synthetic texts with the same satisfaction as human-written ones. This isn't a quality problem on the part of human authors—it's a perception problem on the part of the readers. When the market cannot or will not distinguish between original and synthetic, authenticity becomes a luxury good. The only salvation for human authors: radical personal perspectives that no machine can simulate.